OpenCL: Programming Heterogeneous Architectures

Transcription

1 OpenCL: Programming Heterogeneous Architectures Porting BigDFT to OpenCL Brice Videau (LIG - NANOSIM) April 24, / 45

2 Introduction 2 / 45

3 Needs in computing resources are innite Benets for physicists and chemists More computing power means : Bigger systems, Fewer approximations, Improved accuracy. Numerical experimentation. CEA's hybrid cluster Titane, built by Bull 3 / 45

4 Current and future architectures 3 trends are seen in current calculators : Bigger systems Number of nodes in clusters and grids increase. Number of processors in supercomputer increase. Green Computing Low power components, High eciency, Huge Number of components. More powerful components Increased frequency, Increased number of processors and cores, Specialized co-processors : GPU, CELL... 4 / 45

5 Exploiting those architectures Middlewares are available to program those machines. Each middleware covers a range of usage. Some examples Distributed machines : MPI, KAAPI, CHARM... Multicore architectures : MPI, OpenMP, ompss, StarPU, OpenCL... GPU : OpenCL, NVIDIA Cuda, ATI Streams... 5 / 45

6 Talk Outline 2 OpenCL : a Standard for Parallel Computing 3 Life and Death of OpenCL in a Program 4 Writing Kernels 5 BigDFT 6 Conclusions and perspectives 6 / 45

7 OpenCL : a Standard for Parallel Computing 7 / 45

8 Host-Devices model OpenCL Architecture Model 1 host and several devices. Devices are connected to the host. Host issues commands to the devices. Data transport is done via memory copy. Several devices support OpenCL Host Commands Devices Devices Devices NVIDIA for GPU and in the future for Tegra. AMD and Intel for CPUs and GPUs and MIC? IBM CELL processor. ARM GPUs (Mali) + CPUs 8 / 45

9 Context and Queues Contexts aggregate resources, programs and devices belonging to a common platform (ie NVIDIA, or ATI). Host and devices communicate via buers dened in a context. Commands are sent to devices using command queues. Commands are called kernels. Command queues Can be synchronous or asynchronous. Can be event driven. Several queues can point to the same device, allowing concurrent execution. 9 / 45

10 OpenCL Processing Model Work item 1 Work item 2 Work item 1 Work item 2 Work item n 1 Work item n Work item n 1 Work item n Compute Unit 1 Compute Unit m Compute Device Kernels are split into uni, two or three-dimensional ranges called work groups. Work groups are mapped to compute units. Individual item are processed by work items. 10 / 45

11 OpenCL Memory Model 4 dierent memory space dened on an OpenCL device : Global memory : corresponds to the device RAM, input data are stored there. Constant memory : cached global memory. Local memory : high speed memory shared among work items of a compute unit. Private memory : registers of a work item. 11 / 45

12 Life and Death of OpenCL in a Program The Host Side of OpenCL 12 / 45

13 General Workow Select desired platform Select desired devices Create associated Context Send data to devices using command queues Create command queues associated do devices Compile or load kernels on the devices Send commands to devices using command queues Get data from devices using command queues Release every resources used 13 / 45

14 Platform Selection In a near future every platform will support OpenCL, but the user may not be interested in all of them : select an appropriate platform Get Platforms 1 #i n c l u d e <CL/ c l. h> 2 c l _ u i n t num_platforms ; 3 c l G e t P l a t f o r m I D s ( NULL, NULL, &num_platforms ) ; 4 c l _ p l a t f o r m _ i d p l a t f o r m s = m a l l o c ( s i z e o f ( c l _ p l a t f o r m _ i d ) num_platforms ) ; 5 clgetplatformids ( num_platforms, platforms, NULL ) ; 6 /... / 7 f o r ( i n t i =0; i <num_platforms ; i ++){ 8 /... / 9 c l G e t P l a t f o r m I n f o ( p l a t f o r m s [ i ], CL_PLATFORM_VENDOR,... ) ; 10 /... / 11 } 14 / 45

15 Device Selection Several device from the same vendor is also common : one device for the screen and one device for computations Get Devices 1 #i n c l u d e <CL/ c l. h> 2 c l _ u i n t num_devices ; 3 c l G e t D e v i c e I D s ( p l a t f o r m, CL_DEVICE_TYPE_ALL, NULL, NULL, &num_devices ) ; 4 c l _ d e v i c e _ i d d e v i c e s = m a l l o c ( s i z e o f ( c l _ d e v i c e _ i d ) num_devices ) ; 5 clgetdeviceids ( platform, CL_DEVICE_TYPE_ALL, num_devices, devices, NULL ) ; 6 /... / 7 f o r ( i n t i =0; i <num_devices ; i ++){ 8 /... / 9 c l G e t D e v i c e I n f o ( d e v i c e s [ i ], CL_DEVICE_NAME,... ) ; 10 /... / 11 } 15 / 45

16 Context Creation Context gather devices from the same platform. Those devices will be able to share resources. Create Context 1 c l _ c o n t e x t _ p r o p e r t i e s p r o p e r t i e s [ ] = 2 { CL_CONTEXT_PLATFORM, ( c l _ c o n t e x t _ p r o p e r t i e s ) platform_id, 0 } ; 3 c l _ d e v i c e _ i d d e v i c e s [ ] = { device_id_1, device_id_2 } ; 4 c l _ c o n t e x t c o n t e x t = 5 c l C r e a t e C o n t e x t ( p r o p e r t i e s, 2, d e v i c e s, NULL, NULL, NULL ) ; A shortcut exists, skipping device selection : Create Context from Type 1 c l _ c o n t e x t _ p r o p e r t i e s p r o p e r t i e s [ ] = 2 { CL_CONTEXT_PLATFORM, ( c l _ c o n t e x t _ p r o p e r t i e s ) platform_id, 0 } ; 3 c l _ c o n t e x t c o n t e x t = 4 clcreatecontextfromtype ( p r o p e r t i e s, CL_DEVICE_TYPE_GPU, NULL, NULL, NULL ) ; 16 / 45

17 Building Program from Source Once the context is created, the program is to be built (or loaded from binary). Building Program 1 / s t r i n g s i s an a r r a y o f s t r i n g _ c o u n t NULL t e r m i n a t e d s t r i n g s / 2 cl_program program = 3 clcreateprogramwithsource ( context, string_ count, s t r i n g s, NULL, NULL ) ; 4 / i f d e v i c e _ l i s t i s NULL, program i s b u i l t 5 f o r a l l a v a i l a b l e d e v i c e s i n t h e c o n t e x t / 6 c l B u i l d P r o g r a m ( program, num_devices, d e v i c e _ l i s t, o p t i o n s, NULL, NULL ) ; 7 c l _ k e r n e l k e r n e l = c l C r e a t e K e r n e l ( program, " kernel_name ", NULL ) ; Kernels are extracted from the built program using their name. 17 / 45

18 Creating Command Queues A command queue is used to send commands to a device. They have to be associated with a device. Creating Command Queues 1 cl_command_queue queue = 2 clcreatecommandqueue ( c o n t e x t, d e v i c e s [ chosen_device ], 0, NULL ) ; Options can be specied instead of 0, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE allows for out of order execution for instance. 18 / 45

19 Using OpenCL Using OpenCL is (hopefully) easier than setting it up. Create buffers to store data on devices Send data to devices using command queues Send commands to devices using command queues Get data from devices using command queues 19 / 45

20 Buer Creation In OpenCL buers creation and deletion are explicitly managed. As can be noted buers are tied to a context and not a particular command queue. The implementation is free to transfer buers from devices to host memory or to another device. Creating Simple Buers 1 cl_mem r e a d _ b u f f e r = 2 c l C r e a t e B u f f e r ( c o n t e x t, CL_MEM_READ_ONLY, b u f f e r _ s i z e, NULL, NULL ) ; 3 cl_mem w r i t e _ b u f f e r = 4 c l C r e a t e B u f f e r ( c o n t e x t, CL_MEM_WRITE_ONLY, b u f f e r _ s i z e, NULL, NULL ) ; 20 / 45

21 Pinned Buer Creation Pinned buer creation can oer premium performances. Here is a code sample that can be used on NVIDIA devices. The nals pointers obtained can be used to transfer data between the host and the device. Creating Pinned Simple Buers 1 cl_mem p i n n e d _ r e a d _ b u f f e r = 2 c l C r e a t e B u f f e r ( c o n t e x t, CL_MEM_ALLOC_HOST_PTR CL_MEM_READ_ONLY, 3 b u f f e r _ s i z e, NULL, NULL ) ; 4 cl_mem p i n n e d _ w r i t e _ b u f f e r = 5 c l C r e a t e B u f f e r ( c o n t e x t, CL_MEM_ALLOC_HOST_PTR CL_MEM_WRITE_ONLY, 6 b u f f e r _ s i z e, NULL, NULL ) ; 7 u n s i g n e d c h a r data_in = 8 clenqueuemapbuffer ( queue, pinned_read_buffer, CL_TRUE, CL_MAP_WRITE, 0, 9 b u f f e r _ s i z e, 0, NULL, NULL, NULL ) ; 10 u n s i g n e d c h a r data_out = 11 clenqueuemapbuffer ( queue, pinned_ write_ buffer, CL_TRUE, CL_MAP_READ, 0, 12 b u f f e r _ s i z e, 0, NULL, NULL, NULL ) ; 21 / 45

22 Transferring Data The implementation is free to move buers in memory. But nonetheless, memory is often kept on the device associated to the command queue used to transfer the data. Data Transfer 1 clenqueuewritebuffer ( queue, read_ buffer, CL_TRUE, 0, 2 b u f f e r _ s i z e, data_in, 0, NULL, NULL ) ; 3 / P r o c e s s i n g t h a t r e a d s r e a d _ b u f f e r and w r i t e s w r i t e _ b u f f e r / 4 /... / 5 clenqueuereadbuffer ( queue, write_ buffer, CL_TRUE, 0, 6 b u f f e r _ s i z e, data_out, 0, NULL, NULL ) ; 22 / 45

23 Performing Calculations Once data is transfered, kernels are used to perform calculations. Kernel Usage 1 / P l a c e k e r n e l p a r a m e t e r s i n t h e k e r n e l s t r u c t u r e. / 2 c l S e t K e r n e l A r g ( k e r n e l, 0, s i z e o f ( d a t a _ s i z e ), ( v o i d )& d a t a _ s i z e ) ; 3 c l S e t K e r n e l A r g ( k e r n e l, 1, s i z e o f ( r e a d _ b u f f e r ), ( v o i d )& r e a d _ b u f f e r ) ; 4 c l S e t K e r n e l A r g ( k e r n e l, 2, s i z e o f ( w r i t e _ b u f f e r ), ( v o i d )& w r i t e _ b u f f e r ) ; 5 / Enqueue a 1 d i m e n s i o n a l k e r n e l w i t h a l o c a l s i z e o f 32 / 6 s i z e _ t l o c a l W o r k S i z e [ ] = { 32 } ; 7 s i z e _ t g l o b a l W o r k S i z e [ ] = { shrroundup ( 3 2, d a t a _ s i z e ) } ; 8 clenqueuendrangekernel ( queue, kernel, 1, NULL, 9 globalworksize, localworksize, 0, NULL, NULL ) ; 23 / 45

24 Event Management Almost all functions presented end with : 1..., 0, NULL, NULL ) ; These 3 arguments are used for event management, and thus asynchronous queue handling. Functions can wait for a number of events, and can generate 1 event. 1 event_t e v e n t _ l i s t [ ] = { event1, e v e n t 2 } ; 2 event_t event ; 3 clenqueuereadbuffer ( queue, write_ buffer, CL_FALSE, 0, 4 b u f f e r _ s i z e, data_out, 2, e v e n t _ l i s t, &e v e n t ) ; Previous buer read waits for 2 events and generate a third that will happen when the read is completed. 24 / 45

25 Release Resources OpenCL uses reference counts to manage memory. In order to exit cleanly from an OpenCL program all allocated resources have to be freed : buers (clreleasememobject) events (clreleaseevent) kernel (clreleasekernel) programs (clreleaseprogram) queues (clreleasecommandqueue) context (clreleasecontext) etc / 45

26 Writing Kernels 26 / 45

27 Recall : OpenCL Memory Model 4 dierent memory space dened on an OpenCL device : Global memory : corresponds to the device RAM, input data are stored there. Constant memory : cached global memory. Local memory : high speed memory shared among work items of a compute unit. Private memory : registers of a work item. 27 / 45

28 OpenCL Language : a Subset of C Kernels are written using a C-like language Recursion is prohibited Helper functions are dened Barriers Work item indexes Atomic operations Vector operations New keywords : kernel global, local, constant, private read_only, write_only, read_write 28 / 45

29 Example : Unidimensional Convolutions N e N N l (0,0) (1,0) (0,1) N l N l n N e (0,0) (1,0) (0,1) Input (i,j) Output (i,j) N e One unidimensional convolution with transposition, simple but not too much. Real world code used in BigDFT an electronic structure calculation program. 29 / 45

30 OpenCL memory management Example using a 4*4 block processing a lter of length 5. 0,0 1,0 2,0 3,0 0,1 1,1 2,1 3,1 0,2 1,2 2,2 3,2 0,0 0,1 0,2 0,3 0,3 1,3 2,3 3,3 1,0 1,1 1,2 1,3 0,0 1,0 2,0 3,0 2,0 2,1 2,2 2,3 0,1 1,1 2,1 3,1 3,0 3,1 3,2 3,3 0,2 1,2 2,2 3,2 0,3 1,3 2,3 3,3 1 work item processes one element of the nal matrix. 30 / 45

31 Benets from this approach 3D convolutions can be expressed as a succession of 3 1D convolution/transposition. Memory access are coalesced while reading input matrix and writing output matrix. Bank conicts are almost avoided by padding the buer to the number of bank +1. Maps easily to current architectures. Extends to more complex convolutions found in BigDFT. Almost every convolution in BigDFT have been ported, including free boundary and semi-periodic boundary. 31 / 45

32 Kernel Declaration Kernel Declaration 1 / A c t i v a t e d o u b l e p r e c i s i o n s u p p o r t / 2 #pragma OPENCL EXTENSION cl_khr_fp64 : e n a b l e 3 #d e f i n e FILT_W 16 4 #d e f i n e WG_S 16 5 kernel v o i d 6 attribute ( ( reqd_work_group_size (WG_S,WG_S, 1 ) ) 7 m a g i c f i l t e r 1 d K e r n e l _ d ( u i n t n, u i n t ndat, 8 global c o n s t d o u b l e psi, 9 global d o u b l e out ){ 10 // padded l o c a l b u f f e r s i z e : local d o u b l e tmp [WG_S (WG_S+FILT_W+1)]; Works on double precision oats Kernel expects work group size of 16 x 16 n and ndat are in local memory tmp1 is a storage buer in local memory, shared among work items 32 / 45

33 Work with Indexes Get Indexes and Load Data 1 // g e t o u r p o s i t i o n i n t h e l o c a l work g r o u p 2 c o n s t s i z e _ t i g = g e t _ g l o b a l _ i d ( 0 ) ; 3 c o n s t s i z e _ t j g = g e t _ g l o b a l _ i d ( 1 ) ; 4 // g e t o u r p o s i t i o n i n t h e r e s u l t m a t r i x 5 c o n s t s i z e _ t i = g e t _ l o c a l _ i d ( 0 ) ; 6 c o n s t s i z e _ t j = g e t _ l o c a l _ i d ( 1 ) ; 7 // t r a n s p o s e i n d e x e s i n t h e work g r o u p i n o r d e r t o r e a d t r a n s p o s e d d a t a 8 p t r d i f f _ t i g t = i g i + j FILT_W/ 2 ; 9 p t r d i f f _ t j g t = j g j + i ; 10 // i f we are on the outside, s e l e c t a border element to load, wrapping around 11 //we w i l l be l o a d i n g 2 e l e m e n t s e a c h 12 i f ( i g t < 0 ) 13 tmp [ i (WG_S+FILT_W+1) + j ] = p s i [ j g t + ( n + i g t ) ndat ] ; 14 e l s e 15 tmp [ i (WG_S+FILT_W+1) + j ] = p s i [ j g t + i g t ndat ] ; 16 i g t += FILT_W ; 17 i f ( i g t >= n ) 18 tmp [ i (WG_S+FILT_W+1) + j + FILT_W] = p s i [ j g t + ( i g t n ) ndat ] ; 19 e l s e \n\ 20 tmp [ i (WG_S+FILT_W+1) + j + FILT_W] = p s i [ j g t + i g t ndat ] ; 33 / 45

34 Compute Convolution and Write Output Performing Computations 1 // i n i t i a l i z e r e s u l t 2 d o u b l e t t = 0. 0 ; 3 // r e s t p o s i t i o n i n t h e b u f f e r t o f i r s t e l e m e n t i n v o l v e d i n t h e c o n v o l u t i o n 4 tmp += j 2 (WG_S+FILT_W+1) + i 2 ; 5 // w a i t f o r b u f f e r t o be f u l l 6 b a r r i e r (CLK_LOCAL_MEM_FENCE) ; 7 8 // a p p l y f i l t e r 9 t t += tmp++ FILT0 ; 10 t t += tmp++ FILT1 ; 11 /... / 12 t t += tmp++ FILT15 ; 13 // s t o r e t h e r e s u l t 14 out [ ( j g n+i g )]= t t ; 15 } ; 34 / 45

35 BigDFT OpenCL Port Evaluation 35 / 45

36 Motivations Part of BigDFT was already ported to GPU architectures using CUDA, why a new port? Only part of the program was ported. OpenCL can target several platforms while CUDA can't. OpenCL is a standard, while CUDA is a vendor provided solution. This time most of BigDFT is expected to run on GPU. Code ported is relatively simple and well suited to GPU, thus the performance loss is hoped to be minimal. 36 / 45

37 Objectives Most of BigDFT operations can be expressed as unidimensional convolutions. There are several dozens convolutions to implement. Convolutions of BigDFT share common traits. Objectives Find a common parallelization technique tting most convolutions. Be as ecient as CUDA. 37 / 45

38 Test System Setup GPU 2 : Tesla C2070 (Fermi) 6 GB of RAM Driver version : GPU 2 : Radeon HD GB of RAM Driver version : / 45

39 Test System Setup Host : Lenovo D20 1 Xeon 2.83 GHz (4 Nehalem cores) 8 GB of RAM Linux x86_64 icc / 45

40 Magicfilter Kinetic Synthesis Analysis Kinetic_k Magicfilter_shrink Magicfilter_grow Magicfilter_reverse Compress Uncompress Analysis_shrink Synthesis_grow Introduction OpenCL Basic Management Writing Kernels BigDFT Conclusions and perspectives Comparison CPU,Fermi,HD Performances of CPU vs NVIDIA vs AMD CPU NVIDIA AMD GFLOPS Kernels 40 / 45

41 Magicfilter Kinetic_k Kinetic Magicfilter_grow Synthesis Analysis Magicfilter_shrink Magicfilter_reverse Compress Uncompress Analysis_shrink Synthesis_grow Introduction OpenCL Basic Management Writing Kernels BigDFT Conclusions and perspectives Comparison CPU,Fermi,HD Performances of CPU vs NVIDIA vs AMD CPU NVIDIA AMD Ratio to CPU Kernels 40 / 45

42 Comparison CUDA, OpenCL, CPU 100 Badiane, X Fermi C2070, ZnO 64 at.: CPU vs. Hybrid Percent Speedup 20 0 CPU-mkl CUDA CPU-mkl-mpi OCL-mkl OCL-cublas CUDA-mkl OCL-cublas-mpi CUDA-mkl-mpi CUDA-mpi OCL-mkl-mpi 2 0 Speedup Other CPU Conv LinAlg Comms 41 / 45

43 Hybrid ATI + NVIDIA Tesla C Radeon HD 6970 MPI+NVIDIA/AMD Execution Time (s) Speedup NVIDIA NVIDIA AMD AMD (4 + NV) + (4 + AMD) Table: Performance results for dierent conguration of BigDFT, using MPI + GPUs 42 / 45

44 Conclusions OpenCL OpenCL proved easy to use. Performance is on-par with previous CUDA implementation. Kernels have been shown to run on other architectures : ATI and CPU. BigDFT Full port of BigDFT convolutions on OpenCL. Part of the code was rewritten during this work. Complexity reduced compared to the CUDA support. 43 / 45

45 Perspectives OpenCL Some OpenCL implementations are still recent and buggy. Best way to do multi-gpu, GPU+OpenCL CPU? Optimizing kernels for multiple devices? Automated kernel generation. 44 / 45

46 Questions? Thanks for your attention. 45 / 45